sroot/lgd-cards-gen1

LGD Cards — Gen 1 (dataset-trained) · YOLO11s playing-card detector, 52 classes

A YOLO11s object detector that locates playing cards in an image and names each by rank + suit (52 classes, e.g. 10H, KS, JC). Part of the Live Game Defender (LGD) card-detector family — an on-prem AI integrity monitor for live casino table games. An empty frame yields no detection, which doubles as an "is there a card?" gate.

Gen 1 is the first, dataset-trained generation. It is superseded — for real casino/table footage use lgd-cards-gen3 (current). Gen 1 scores near-perfect on its own dataset's validation split but has a real-world synthetic→real gap (see Metrics). It is published for provenance and reproducibility.

Generations

Gen	Repo	Trained on	Frozen real-video holdout recall	Status
1	lgd-cards-gen1 (this)	Roboflow ow27d v4 dataset	— (dataset-val only)	superseded
2	lgd-cards-gen2	+ day-1 PoC table video	0.68	superseded
3	lgd-cards-gen3	+ day-2 PoC table video	0.85	✅ current

Chip detectors: lgd-chips-gen1 · lgd-chips-gen2.

Classes (52, in data.yaml order)

10C 10D 10H 10S 2C 2D 2H 2S 3C 3D 3H 3S 4C 4D 4H 4S 5C 5D 5H 5S
6C 6D 6H 6S 7C 7D 7H 7S 8C 8D 8H 8S 9C 9D 9H 9S
AC AD AH AS JC JD JH JS KC KD KH KS QC QD QH QS

Suit codes: C=Clubs, D=Diamonds, H=Hearts, S=Spades.

Files

• best.pt — Ultralytics PyTorch weights.
• best.onnx — ONNX export (run with onnxruntime, no training framework needed).
• data.yaml — class list / dataset config.

Training

• Base: Ultralytics yolo11s.pt (COCO-pretrained), fine-tuned 25 epochs, imgsz=640.
• Data: Roboflow augmented-startups/playing-cards-ow27d v4 — Public Domain; 21,203 train / 2,020 val / 1,010 test.
• Hardware: NVIDIA RTX 3060 (12 GB).

Metrics — on the dataset's own validation split

mAP50	mAP50-95	precision	recall
0.995	~0.80	0.999	~1.00

⚠️ These are dataset-distribution numbers, NOT real-world casino accuracy. On real webcam photos this generation generalizes to many cards but has a known synthetic→real gap on some black court cards (e.g. King of Spades) — where it tends to return no detection rather than a wrong guess. This is exactly why gens 2–3 were fine-tuned on real table video. Do not treat the table above as real-table accuracy.

Usage

from ultralytics import YOLO
model = YOLO("best.pt")
r = model.predict("frame.jpg", conf=0.25)[0]
for b in r.boxes:
    print(r.names[int(b.cls)], float(b.conf))

# ONNX / onnxruntime (MIT) — no AGPL code on the inference path
import onnxruntime as ort
sess = ort.InferenceSession("best.onnx", providers=["CPUExecutionProvider"])
# 640x640 letterboxed input; class order == data.yaml above

License & provenance

AGPL-3.0. Every model in this family is a fine-tune of Ultralytics YOLO11 (yolo11s.pt), which is AGPL-3.0 — so these weights inherit AGPL-3.0 and are not an original work of ours. If you deploy them in a networked service, AGPL §13 applies: you must offer users the Corresponding Source. Running the ONNX export via onnxruntime keeps the inference code AGPL-free, but the AGPL still attaches to the weights themselves.

Built for Live Game Defender (LGD) — an on-prem AI integrity monitor for live casino table games. © 2026 TechTools s.r.o.